Direct Asset Prediction

This family covers models that predict asset-level signals directly rather than first estimating latent structure.

SAEModel

SAEModel means supervised autoencoder in this library.

It does not mean:

• sparse autoencoder
• unsupervised autoencoder
• latent-factor model

The current implementation follows the supervised autoencoder pattern used in the Jane Street competition lineage:

• encoder / bottleneck
• decoder
• auxiliary head
• main predictive head

The key architectural idea is not "autoencoding for its own sake." The bottleneck and decoder regularize the representation that the predictive heads use. The reconstruction task is there to improve the supervised signal, not to recover a structural latent-factor decomposition.

Why It Lives Outside Latent Factors

The library originally explored treating SAE as another latent-factor model. That turns out to be the wrong abstraction.

In the current design:

• SAEModel is a direct predictor
• it consumes CrossSectionBatch
• it emits AssetSignalResult

So the workflow is:

cross-section batch -> supervised autoencoder -> asset signals

not:

cross-section batch -> factor state -> factor forecaster -> beta × lambda

Checkpoints

SAEModel supports:

• checkpoint_interval
• checkpoint_epochs
• default_checkpoint

so you can evaluate intermediate training horizons explicitly.

Example

from ml4t.models import CrossSectionBatch, SAEConfig, SAEModel

model = SAEModel(SAEConfig(n_epochs=50, checkpoint_interval=5))
fit_summary = model.fit(train_batch, validation_batch=val_batch)
signals = model.predict(test_batch)

Outputs

predict() returns:

• AssetSignalResult

with:

• signal_values
• timestamps
• asset IDs
• selected checkpoint metadata

When To Use It

Use SAEModel when you want:

• a direct supervised predictor
• a strong tabular deep-learning baseline for cross-sectional signals
• checkpoint-aware asset-level predictions

This is the right family when the modeling question is:

• "Can I learn a useful cross-sectional signal directly?"

Use latent-factor models instead when the question is:

• "Can I explain returns through exposures to a small latent factor system?"